GitHub Copilot's Autopilot Mode: Fully Autonomous Coding in VS Code

GitHub dropped its biggest Copilot update in months on April 8, 2026 — a combined changelog for VS Code versions 1.111 through 1.115. The headline addition is Autopilot mode: agents can now approve their own tool calls, automatically retry on errors, and keep working until a task is done without asking for your permission at every step.

This is a direct answer to Cursor 3.0's parallel agents — and it lives inside VS Code, which is already on most developers' machines. You don't need a separate tool.

How Autopilot Mode Works

Before Autopilot, Copilot agents had to ask for your approval before running terminal commands, writing files, or calling external tools. Useful for safety, but it created friction on long multi-step tasks. You'd start a task and then sit clicking 'Approve' every few minutes.

Autopilot removes that loop. Once you enable it, the agent proceeds through its entire task plan — reading files, writing code, running tests, installing packages, fixing errors — without pausing. You get the result when it's done.

How to enable Autopilot mode in VS Code:
1. Open GitHub Copilot Chat (Ctrl+Shift+I or Cmd+Shift+I)
2. Click the gear icon next to the model selector
3. Find 'Agent Approval Mode' in settings
4. Choose from three levels:
   - Default: agent asks before running terminal commands and file writes
   - Bypass Approvals: agent auto-approves safe actions, asks for risky ones
   - Autopilot: fully autonomous — agent approves everything and retries errors
5. You can also set this in settings.json: "github.copilot.agent.approvalMode": "autopilot"

Nested Subagents: Agents That Spawn Agents

The April 8 update also shipped nested subagents. Your main Copilot agent can now spawn child agents to handle subtasks, then collect their results. This enables much more complex automated workflows than a single agent can run in sequence.

Example: an agent tasked with 'add user authentication' might spawn one subagent to research the existing session handling, one to write the database schema changes, and one to write the UI components — all in parallel — then synthesize the results into a coherent implementation.

Configurable Thinking Effort

You can now dial how deeply reasoning models think before responding. The setting appears directly in the model picker and persists across conversations. For quick autocomplete and simple questions, set thinking effort low. For architecture planning or debugging complex issues, set it high.

When to use high thinking effort:
- Debugging a non-obvious bug that involves multiple interacting systems
- Planning a large refactor before writing any code
- Designing an API structure that other teams will depend on
- Security-sensitive code where missing an edge case matters

When to use low thinking effort:
- Writing boilerplate you'll review immediately anyway
- Simple one-file changes
- Explaining what existing code does
- Quick syntax fixes

Video in Chat: Attach Screen Recordings

You can now attach screenshots and video recordings to Copilot Chat messages. When an agent makes changes, it can return a screen recording showing what it built. This sounds minor but is genuinely useful for frontend work — instead of describing 'the button is misaligned,' you can record the bug and send the video directly to Copilot Chat.

Integrated Browser Debugging

VS Code now includes built-in browser debugging (new editor-browser debug type). Set breakpoints, step through JavaScript, and inspect variables in a browser-like environment without leaving VS Code. Combined with Autopilot, this means agents can debug frontend behavior the same way they debug backend behavior.

What to Do Right Now

Update VS Code to 1.115 if you haven't already. Autopilot mode and nested subagents are in public preview — they're available now but may still have rough edges. The safest way to explore Autopilot is on a feature branch with a clear scope: pick a contained task, set Autopilot, and watch what it does before trusting it on anything in main.